Objective 1:
The project built proof of principle implementations of feed forward quantum neural networks. For a quantum neural network based on adiabatic ramp activation, WMI has implemented the activation protocol on a two-qubit chip and measured the sigmoid-type activation function on the target qubit and has thus conducted a proof-of-principle experiment demonstrating the adiabatic process.
In the search for first applications, where quantum neural networks can outperform classical artificial intelligence, we have identified the analysis of many-qubit quantum states as a very promising candidate. Such quantum states will be produced as outputs of near-term quantum hardware, making this direction very promising for future developments. ETH Zürich and FAU have made a first step in this promising direction by demonstrating a quantum convolutional neural network that is able to classify quantum phases of quantum matter. FAU has explored generalizations of quantum convolutional neural networks to other systems and larger depth, to show that these networks indeed form a powerful tool for the analysis and classification of quantum many-body states.
Objective 2:
TU Delft has made progress towards realizing a non-equilibrium quantum annealer. The analysis so far clearly indicates that building the device is feasible, and its operation are realizable.
Objective 3:
FAU and ETH Zürich have developed a novel convolutional quantum neural network, that reduces the number of measurements required for reliable, high accuracy recognition of quantum phases by a factor that grows exponentially in the system size. They also showed that the number of required quantum gates can be reduced to one layer of controlled phase gates only, thus significantly reducing the needed circuit depth.
To shorten such circuits, ETHZ has implemented variable angle two-qubit gates directly in hardware. To address the challenge of required network connectivity, FAU and WMI have developed a scheme to activate a quantum neuron controlled by the states of two quantum neurons in the preceding layer. This scheme also forms a quantum Toffoli gate. UPV/EHU have developed a data-reuploading quantum neural network that is found to show better approximation capabilities than classical neural networks for approximating continuous functions and thus may allow to achieve better performance with less resources.
Objective 4:
Via numerical simulations of various types of quantum neural networks, FAU has conducted a comparative study of proposed approaches when applied to a classification problem for real world data and when applied to the pure quantum task of learning a unitary.
VW has designed a variational quantum algorithm approach to train a quantum learning (Q-learning) agent for a reinforcement learning problem. Differently to the conventional policy gradient approach, the agent here learns to approximate the future expected reward for a given state and action pair based on a given policy. The concept has been developed further by FAU by introducing more powerful action decoding procedures, i.e. mappings between the readout data of the quantum neural network and the decision for an action of the agent.
